---
license: apache-2.0
base_model: DataoceanAI/dolphin-small
library_name: onnx-asr
pipeline_tag: automatic-speech-recognition
language:
- ar
- az
- ba
- bn
- fa
- fil
- gu
- hi
- id
- ja
- jv
- kab
- kk
- km
- ko
- ks
- ky
- lo
- mn
- mr
- ms
- my
- ne
- or
- pa
- ps
- ru
- si
- su
- ta
- te
- tg
- th
- tl
- ug
- ur
- uz
- vi
- yue
- zh
tags:
- automatic-speech-recognition
- onnx
- espnet
- e-branchformer
- dolphin
---

# dolphin-small-onnx

ONNX export of [DataoceanAI/dolphin-small](https://huggingface.co/DataoceanAI/dolphin-small),
a 372M parameter E-Branchformer encoder (12 blocks, width 768) with a 12 block transformer
decoder, trained by **DataoceanAI** on 40 Eastern languages and 22 Chinese dialects.
Apache-2.0, the same licence as the source model. All credit for the model goes to
DataoceanAI; this repository only holds the converted graphs.

## Files

| File | Size |
| --- | --- |
| `encoder.onnx` + `.data` | 700 MB |
| `decoder.onnx` + `.data` | 716 MB |
| `encoder.int8.onnx` + `.data` | 219 MB |
| `decoder.int8.onnx` + `.data` | 283 MB |

`encoder.onnx` takes the raw 16 kHz waveform: the ESPnet `default` frontend (STFT
512/400/160 plus 80 log-mel) and the `global_mvn` statistics are part of the graph, so
`config.json` declares `"preprocessor": "identity"`.

## Usage

Needs the `feat/espnet-aed-prompts` branch of the
[TigreGotico/onnx-asr](https://github.com/TigreGotico/onnx-asr) fork, which adds the
config-driven decode prompt that this model needs.

```sh
pip install "onnx-asr[cpu,hub] @ git+https://github.com/TigreGotico/onnx-asr@feat/espnet-aed-prompts"
```

```py
import onnx_asr

model = onnx_asr.load_model("OpenVoiceOS/dolphin-small-onnx")
print(model.recognize("speech.wav", language="ja"))

# Chinese dialects use the full tag.
print(model.recognize("speech.wav", language="zh-SICHUAN"))

# int8
model = onnx_asr.load_model("OpenVoiceOS/dolphin-small-onnx", quantization="int8")
```

The `language` argument accepts a full tag (`zh-CN`), the same tag with an underscore
(`zh_CN`) or a bare language (`zh`), which maps to the first region of that language.
Leave `language` out and the model predicts the language and the region itself.

## Languages

`zh-CN`, `zh-TW`, `zh-WU`, `zh-SICHUAN`, `zh-SHANXI`, `zh-ANHUI`, `zh-TIANJIN`, `zh-NINGXIA`, `zh-SHAANXI`, `zh-HEBEI`, `zh-SHANDONG`, `zh-GUANGDONG`, `zh-SHANGHAI`, `zh-HUBEI`, `zh-LIAONING`, `zh-GANSU`, `zh-FUJIAN`, `zh-HUNAN`, `zh-HENAN`, `zh-YUNNAN`, `zh-MINNAN`, `zh-WENZHOU`, `ja-JP`, `th-TH`, `ru-RU`, `ko-KR`, `id-ID`, `vi-VN`, `ct-NULL`, `ct-HK`, `ct-GZ`, `hi-IN`, `ur-IN`, `ur-PK`, `ms-MY`, `uz-UZ`, `ar-MA`, `ar-GLA`, `ar-SA`, `ar-EG`, `ar-KW`, `ar-LY`, `ar-JO`, `ar-AE`, `ar-LVT`, `fa-IR`, `bn-BD`, `ta-SG`, `ta-LK`, `ta-IN`, `ta-MY`, `te-IN`, `ug-NULL`, `ug-CN`, `gu-IN`, `my-MM`, `tl-PH`, `kk-KZ`, `or-IN`, `ne-NP`, `mn-MN`, `km-KH`, `jv-ID`, `lo-LA`, `si-LK`, `fil-PH`, `ps-AF`, `pa-IN`, `kab-NULL`, `ba-NULL`, `ks-IN`, `tg-TJ`, `su-ID`, `mr-IN`, `ky-KG`, `az-AZ`

## Accuracy

Greedy decoding, four FLEURS test clips, against the native Dolphin implementation with
the same greedy decoding and the same forced language and region.

| Clip | fp32 | int8 |
| --- | --- | --- |
| zh (cmn_hans_cn) | character-exact | character-exact |
| ja (ja_jp) | character-exact | one word differs |
| ko (ko_kr) | character-exact | character-exact |
| th (th_th) | character-exact | one word differs, more word spacing |

Real-time factor on 12 idle cores of an AMD box, one clip at a time: **0.24 fp32**,
**0.13 int8**.

## Limitations

* Greedy decoding only. The native implementation defaults to beam search.
* No timestamps. The `<notimestamp>` token is part of the baked prompt.
* Hotword biasing of the source model is not exported.